[bri-prism]

Fixes #241.

Problem

release-prism.yml builds the Windows CUDA x64, Linux CUDA and Windows HIP bundles with GGML_NATIVE=OFF but without GGML_BACKEND_DL / GGML_CPU_ALL_VARIANTS. With native detection off and no explicit ISA flags, ggml turns on AVX/AVX2/FMA/F16C by default. Each of those bundles therefore ships a single AVX2-only CPU backend, and on a CPU without AVX2 the process dies at startup with no message. In #241, even llama-server --version exits silently.

This is confirmed in the shipped prism-b10709-9a9394a artifacts:

bundle	CPU backend shipped	AVX2/FMA code present
win-cpu-x64	15 variants (ggml-cpu-x64 … -zen4)	only in the AVX2+ variants
win-vulkan-x64	9 variants (fixed in #198)	only in the AVX2+ variants
win-cuda-13.3-x64	one ggml-cpu.dll	yes: 11,734 ymm ops, 458 vfmadd, 91 AVX2-only permutes/broadcasts
win-hip-radeon-x64	one ggml-cpu.dll	yes: 1,571 vfmadd, 2,004 AVX2-only ops
linux-cuda-*-x64	one libggml-cpu.so	(same config gap)

For comparison, ggml-cpu-x64.dll and ggml-cpu-sse42.dll in the CPU zip contain 0 vfmadd and 0 AVX2-only ops.

Fix

Add -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON to those three jobs, as the CPU and Vulkan bundles already do. The arm64 CUDA leg already used BACKEND_DL with GGML_CPU=OFF and is unchanged.

Verification (local, Windows)

I have no CUDA toolkit on this machine, so I verified the CPU side of the Windows CUDA job's configuration with the same flags and the same shared CMAKE_ARGS, minus GGML_CUDA. MSVC 19.44, -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON, on prism @ 842b18804:

• All nine MSVC CPU variants build: x64, sse42, sandybridge, haswell, alderlake, skylakex, cannonlake, cascadelake, icelake. They also build with LLAMA_FATAL_WARNINGS=ON (/WX), so the non-AVX2 x86 tiers that broke in ggml-cpu : fix PQ2_0 x86 vec_dot build on non-AVX2 CPU variants #263 are clean under MSVC now. With /WX the only failures are pre-existing and unrelated: src/llama.cpp C4297 and an STL C4244. None of these jobs set fatal warnings.
• The full bundle builds (474/474), and llama-server --version runs.
• Runtime variant selection works. A CPU-only llama-server on a Core Ultra X7 358H (AVX2 + AVX-VNNI) maps ggml-cpu-alderlake.dll. A CPU without AVX2 takes the same scoring path down to sse42/x64, as the CPU zip already does for the Misc. bug: not start on Windows with CUDA #241 reporter ("only CPU version … work right").

Not verified here: the CUDA and HIP legs themselves (no toolkit on this box), and the Linux CUDA leg's GCC variant build. A CI run of release-prism.yml on this branch is the real check.

🤖 Generated with Claude Code

https://claude.ai/code/session_012Q6ecxozUXFFFLE3N2SMiR

[Copilot]

Copilot review overview

🟢 Approval recommended

The functional changes are consistent with existing release jobs; only a minor comment-format issue remains.

Review effort: Balanced
Findings: 1

Open (1)

• Collapse verbose comment into one concise line · New

What changed in this PR

Adds runtime-selected CPU backend variants to CUDA and HIP release bundles, preventing startup failures on non-AVX2 CPUs.

Changes:

• Enables dynamic CPU backend loading and all CPU variants for Linux CUDA, Windows CUDA x64, and Windows HIP.
• Preserves the Windows CUDA ARM64 GPU-only behavior.

File	Description
.github/​workflows/​release-prism.yml	Enables compatible CPU variants in GPU release bundles.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.